{
  "id": 318445,
  "title": "Low accuracy on TPU",
  "url": "/competitions/ultra-mnist/discussion/318445",
  "author_name": "Shankar Mahadevan",
  "post_date": "2022-04-12T10:50:54.780000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I am trying a few approaches for the innovative part, and getting very low accuracy when training on TPUs. The same approach gives a comparatively better result in GPU. I am using a large batch size (~512) and also SGD (Since it is preferred in TPUs, I have tried other optimizers too) as the optimizer. I have previously had success in training models with TPUs (Transformers, CNNs, etc), but in this competition, I'm not able to get good accuracy. What may be the cause of this?</p>",
  "messages": [
    {
      "id": 1752965,
      "postDate": "2022-04-12T10:50:54.780Z",
      "content": "<p>I am trying a few approaches for the innovative part, and getting very low accuracy when training on TPUs. The same approach gives a comparatively better result in GPU. I am using a large batch size (~512) and also SGD (Since it is preferred in TPUs, I have tried other optimizers too) as the optimizer. I have previously had success in training models with TPUs (Transformers, CNNs, etc), but in this competition, I'm not able to get good accuracy. What may be the cause of this?</p>",
      "rawMarkdown": "I am trying a few approaches for the innovative part, and getting very low accuracy when training on TPUs. The same approach gives a comparatively better result in GPU. I am using a large batch size (~512) and also SGD (Since it is preferred in TPUs, I have tried other optimizers too) as the optimizer. I have previously had success in training models with TPUs (Transformers, CNNs, etc), but in this competition, I'm not able to get good accuracy. What may be the cause of this?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1752965": "I am trying a few approaches for the innovative part, and getting very low accuracy when training on TPUs. The same approach gives a comparatively better result in GPU. I am using a large batch size (~512) and also SGD (Since it is preferred in TPUs, I have tried other optimizers too) as the optimizer. I have previously had success in training models with TPUs (Transformers, CNNs, etc), but in this competition, I'm not able to get good accuracy. What may be the cause of this?"
  }
}